RINGS: REALTIME: Resilient Edge-cloud Autonomous Learning with Timely Inferences

Rutgers University New Brunswick, New Brunswick NJ

Investigators

Abstract

Machine learning (ML) is the enabler of emerging real-time applications ranging from augmented reality and smart cities to autonomous vehicles that are changing how people live and work. Low latency is essential for these services; emerging real-time applications will typically need assistance from a mobile edge cloud (MEC) for real-time operation. This emerging scenario introduces significant new challenges: mobile devices are heterogeneous, ranging from energy-harvesting sensors to automobiles, but storage and compute resources are generally limited and communication is often over low-bandwidth channels; real-time deployment of trained ML models requires autonomous computation and decision-making that is adaptive to heterogeneous time-varying local environments; devices need to make high-accuracy inferences on high-dimensional data in real time; devices continuously gather new data that must be processed, aggregated, and communicated to the MEC; mobile users have heterogenous privacy preferences that require privacy-sensitive use of the MEC; and the applications and services on the mobile devices must be resilient to changes in both the cyber and physical worlds in order to ensure personal safety. This project is aimed at the design and experimental validation of an MEC-based distributed ML system that accounts for these factors. In this setting of real-time operation, online decision-making, and offline training of ML-based applications that must be resilient to data, application, user, and system changes, this research program has four facets: (1) Edge-centric distributed ML models to enable both real-time inferences at mobile devices and fast distributed semi-supervised training are being developed and evaluated. (2) Based on age-of-information timeliness metrics, real-time inference methods and system operation are optimized to balance mobile computation against network resources. (3) Differential privacy and other privacy metrics for real-time and online operation of MEC-assisted ML are being developed and incorporated in the distributed algorithms for system adaptation. (4) The project integrates these design approaches in a proof-of-concept prototype on the NSF COSMOS testbed in NY City to validate feasibility and evaluate device and system resilience for representative applications. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.